Official platform of Tsinghua Big Data Research Center, sharing the team's latest research, teaching updates, and big data news.
The article dissects DeepSeek's MoE model‑parallel strategy, explaining how GPU compute and communication are overlapped through expert, pipeline, and ZeRO‑1 parallelism, and introduces DualPipe and Waved‑EP kernels that enable efficient training on large‑scale hardware.
The article explains how AI agents use memory to retain conversation context, why sending the full history to large language models is inefficient, and presents rolling context windows, inverted‑index pruning, semantic embedding retrieval, and GraphRAG as complementary strategies to build more accurate and personalized agents.
This article systematically analyzes the most influential open‑source foundation models—Meta Llama, Alibaba Qwen, Mistral AI, and others—detailing their core architectures, lightweight, instruction‑tuned, multimodal, and industry‑specific derivatives, and outlining current ecosystem characteristics and future development trends.
The article analyzes SubQ, a new LLM architecture using Subquadratic Sparse Attention (SSA) to achieve a 12‑million‑token context window with linear compute scaling, delivering up to 52× speedup and costing just 5% of Opus while matching dense‑attention performance on long‑context benchmarks.
The StreakMind system built by the Spanish Royal Navy Academy uses a YOLO11‑OBB detector trained on over 2,000 real astronomical images and 280 synthetic streaks to automatically identify satellite and asteroid streaks with 94% precision and 97% recall, delivering standardized database entries and robust frame‑to‑frame tracking.
The 2026 China University Computer Competition – Big Data Challenge reveals the Monthly Star award winners, each receiving 800 RMB, and presents detailed experience reports from the top teams covering feature engineering, model selection, training validation, and ensemble strategies for stock prediction.
TimechoAI, the cloud service built on the Timer time‑series large model, offers the latest SOTA model (Timer‑3.5) alongside classic baselines, supports multiple data input methods, covariate integration, and API/SDK access, and invites industrial and IoT teams to test its predictive maintenance, production optimization, energy load forecasting, and anomaly detection capabilities through a simple invitation process.
This guide explains what Claude Code skills are, categorizes common skill types, provides concrete examples for each category, and offers detailed best‑practice recommendations for building, testing, sharing, and managing skills within Claude’s AI ecosystem.
MathForge tackles the long‑standing question of which math problems deserve focus in reinforcement‑learning‑based training, introducing a difficulty‑aware optimizer (DGPO) and multi‑aspect question reformulation (MQR) that together prioritize harder‑but‑learnable questions, yielding consistent performance gains across model sizes and modalities.
The article introduces time‑series data, its ubiquitous examples, the challenges of traditional small models, and proposes a universal time‑series large model that simplifies data preparation and model building, ultimately enabling more efficient and stable industrial AI solutions, now offered as a cloud service.
